README 10.8 KB
Newer Older
Iustin Pop's avatar
Iustin Pop committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
Cluster tools (h-aneti?)
========================

These are some simple cluster tools for fixing common problems. Right now N+1
and rebalancing are included.

.. contents::

Cluster N+1 solver
------------------

This program runs a very simple brute force algorithm over the instance
placement space in order to determine the shortest number of replace-disks
needed to fix the cluster. Note this means we won't get a balanced cluster,
just one that passes N+1 checks.

Also note that the set of all instance placements on a 20/80 cluster is
(20*19)^80, that is ~10^200, so...

Algorithm
+++++++++

The algorithm is a simple two-phase process.

In phase 1 we determine the removal set, that is the set of instances that when
removed completely from the cluster, make it healthy again. The instances that
can go into the set are all the primary and secondary instances of the failing
nodes. The result from this phase is actually a list - we compute all sets of
the same minimum length.

So basically we aim to determine here: what is the minimum number of instances
that need to be removed (this is called the removal depth) and which are the
Iustin Pop's avatar
Iustin Pop committed
33
actual combinations that fit (called the list of removal sets).
Iustin Pop's avatar
Iustin Pop committed
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230

In phase 2, for each removal set computed in the previous phase, we take the
removed instances and try to determine where we can put them so that the
cluster is still passing N+1 checks. From this list of possible solutions
(called the list of solutions), we compute the one that has the smallest delta
from the original state (the delta is the number of replace disks that needs to
be run) and chose this as the final solution.

Implementation
++++++++++++++

Of course, a naive implementation based on the above description will run for
long periods of time, so the implementation has to be smart in order to prune
the solution space as eagerly as possible.

In the following, we use as example a set of test data (a cluster with 20
nodes, 80 instances that has 5 nodes failing N+1 checks for a total of 12
warnings).

On this set, the minimum depth is 4 (anything below fails), and for this depth
the current version of the algorithm generates 5 removal sets; a previous
version of the first phase generated a slightly different set of instances, with
two removal sets. For the original version of the algorithm:

- the first, non-optimized implementation computed a solution of delta=4 in 30
  minutes on server-class CPUs and was still running when aborted 10 minutes
  later
- the intermediate optimized version computed the whole solution space and
  found a delta=3 solution in around 10 seconds on a laptop-class CPU (total
  number of solutions ~600k)
- latest version on server CPUs (which actually computes more removal sets)
  computes depth=4 in less than a second and depth=5 in around 2 seconds, and
  depth=6 in less than 20 seconds; depth=8 takes under five minutes (this is
  10^10 bigger solution space)

Note that when (artificially) increasing the depth to 5 the number of removal
sets grows fast (~3000) and a (again artificial) depth 6 generates 61k removal
sets. Therefore, it is possible to restrict the number of solution sets
examined via a command-line option.

The factors that influence the run time are:

- the removal depth; for each increase with one of the depth, we grow the
  solution space by the number of nodes squared (since a new instance can live
  any two nodes as primary/secondary, therefore (almost) N times N); i.e.,
  depth=1 will create a N^2 solution space, depth two will make this N^4,
  depth three will be N^6, etc.
- the removal depth again; for each increase in the depth, there will be more
  valid removal sets, and the space of solutions increases linearly with the
  number of removal sets

Therefore, the smaller the depth the faster the algorithm will be; it doesn't
seem like this algorithm will work for clusters of 100 nodes and many many
small instances (e.g. 256MB instances on 16GB nodes).

Currently applied optimizations:

- when choosing where to place an instance in phase two, there are N*(N-1)
  possible primary/secondary options; however, if instead of iterating over all
  p * s pairs, we first determine the set of primary nodes that can hold this
  instance (without failing N+1), we can cut (N-1) secondary placements for
  each primary node removed; and since this applies at every iteration of phase
  2 it linearly decreases the solution space, and on full clusters, this can
  mean a four-five times reductions of solution space
- since the number of solutions is very high even for smaller depths (on the
  test data, depth=4 results in 1.8M solutions) we can't compare them at the
  end, so at each iteration in phase 2 we only promote the best solution out of
  our own set of solutions
- since the placement of instances can only increase the delta of the solution
  (placing a new instance will add zero or more replace-disks steps), it means
  the delta will only increase while recursing during phase 2; therefore, if we
  know at one point that we have a current delta that is equal or higher to the
  delta of the best solution so far, we can abort the recursion; this cuts a
  tremendous number of branches; further promotion of the best solution from
  one removal set to another can cut entire removal sets after a few recursions

Command line usage
++++++++++++++++++

Synopsis::

    hn1 [-n NODES_FILE] [-i INSTANCES_FILE] [-d START_DEPTH] \
        [-r MAX_REMOVALS] [-l MIN_DELTA] [-L MAX_DELTA] \
        [-p] [-C]

The -n and -i options change the names of the input files. The -d option
changes the start depth, as a higher depth can give (with a longer computation
time) a solution with better delta. The -r option restricts at each depth the
number of solutions considered - with r=1000 for example even depth=10 finishes
in less than a second.

The -p option will show the cluster state after the solution is implemented,
while the -C option will show the needed gnt-instance commands to implement
it.

The -l (--min-delta) and -L (--max-delta) options restrict the solution in the
following ways:

- min-delta will cause the search to abort early once we find a solution with
  delta less than or equal to this parameter; this can cause extremely fast
  results in case a desired solution is found quickly; the default value for
  this parameter is zero, so once we find a "perfect" solution we finish early
- max-delta causes rejection of valid solution but which have delta higher
  than the value of this parameter; this can reduce the depth of the search
  tree, with sometimes significant speedups; by default, this optimization is
  not used

Individually or combined, these two parameters can (if there are any) very
fast result; on our test data, depth=34 (max depth!) is solved in 2 seconds
with min-delta=0/max-delta=1 (since there is such a solution), and the
extremely low max-delta causes extreme pruning.

Cluster rebalancer
------------------

Compared to the N+1 solver, the rebalancer uses a very simple algorithm:
repeatedly try to move each instance one step, so that the cluster score
becomes better. We stop when no further move can improve the score.

The algorithm is divided into rounds (all identical):

#. Repeat for each instance:

    #. Compute score after the potential failover of the instance

    #. For each node that is different from the current primary/secondary

        #. Compute score after replacing the primary with this new node

        #. Compute score after replacing the secondary with this new node


    #. Out of this N*2+1 possible new scores (and their associated move) for
       this instance, we choose the one that is the best in terms of cluster
       score, and then proceed to the next instance

Since we don't compute all combinations of moves for instances (e.g. the first
instance's all moves Cartesian product with second instance's all moves, etc.)
but we proceed serially instance A, then B, then C, the total computations we
make in one steps is simply N(number of nodes)*2+1 times I(number of instances),
instead of (N*2+1)^I. So therefore the runtime for a round is trivial.

Further rounds are done, since the relocation of instances might offer better
places for instances which we didn't move, or simply didn't move to the best
place. It is possible to limit the rounds, but usually the algorithm finishes
after a few rounds by itself.

Note that the cluster *must* be N+1 compliant before this algorithm is run, and
will stay at each move N+1 compliant. Therefore, the final cluster will be N+1
compliant.

Single-round solutions
++++++++++++++++++++++

Single-round solutions have the very nice property that they are
incrementally-valid. In other words, if you have a 10-step solution, at each
step the cluster is both N+1 compliant and better than the previous step.

This means that you can stop at any point and you will have a better cluster.
For this reason, single-round solutions are recommended in the common case of
let's make this better. Multi-round solutions will be better though when adding
a couple of new, empty nodes to the cluster due to the many relocations needed.


Multi-round solutions
+++++++++++++++++++++

A multi-round solution (not for a single round), due to de-duplication of moves
(i.e. just put the instance directly in its final place, and not move it five
times around) loses both these properties. It might be that it's not possible to
directly put the instance on the final nodes. So it can be possible that yes,
the cluster is happy in the final solution and nice, but you cannot do the steps
in the shown order. Solving this (via additional instance move(s)) is left to
the user.

Command line usage
++++++++++++++++++

Synopsis::

    hbal [-n NODES_FILE] [-i INSTANCES_FILE] \
         [-r MAX_ROUNDS] \
         [-p] [-C]

The -n and -i options change the names of the input files. The -r option
restricts the maximum number of rounds (and is more of safety measure).

The -p option will show the cluster state after the solution is implemented,
while the -C option will show the needed gnt-instance commands to implement
it.

Integration with Ganeti
-----------------------

The programs needs only the output of the node list and instance list. That is,
they need the following two commands to be run::

Iustin Pop's avatar
Iustin Pop committed
231
    gnt-node list -oname,mtotal,mfree,dtotal,dfree \
Iustin Pop's avatar
Iustin Pop committed
232
      --separator '|' --no-headers > nodes
Iustin Pop's avatar
Iustin Pop committed
233
    gnt-instance list -oname,admin_ram,sda_size,pnode,snodes \
Iustin Pop's avatar
Iustin Pop committed
234
235
236
237
238
239
240
241
242
243
244
245
      --separator '|' --no-head > instances

These two files should be saved under the names of 'nodes' and 'instances'.

When run, the programs will show some informational messages and output the
chosen solution, in the form of a list of instance name and chosen
primary/secondary nodes. The user then needs to run the necessary commands to
get the instances to live on those nodes.

Note that sda_size is less than the total disk size of an instance by 4352
MiB, so if disk space is at a premium the calculation could be wrong; in this
case, please adjust the values manually.